Make sure we copy pages inserted with "vm_insert_page()" on fork
[linux-2.6.git] / mm / mremap.c
1 /*
2  *      mm/mremap.c
3  *
4  *      (C) Copyright 1996 Linus Torvalds
5  *
6  *      Address space accounting code   <alan@redhat.com>
7  *      (C) Copyright 2002 Red Hat Inc, All Rights Reserved
8  */
9
10 #include <linux/mm.h>
11 #include <linux/hugetlb.h>
12 #include <linux/slab.h>
13 #include <linux/shm.h>
14 #include <linux/mman.h>
15 #include <linux/swap.h>
16 #include <linux/fs.h>
17 #include <linux/highmem.h>
18 #include <linux/security.h>
19 #include <linux/syscalls.h>
20
21 #include <asm/uaccess.h>
22 #include <asm/cacheflush.h>
23 #include <asm/tlbflush.h>
24
25 static pmd_t *get_old_pmd(struct mm_struct *mm, unsigned long addr)
26 {
27         pgd_t *pgd;
28         pud_t *pud;
29         pmd_t *pmd;
30
31         pgd = pgd_offset(mm, addr);
32         if (pgd_none_or_clear_bad(pgd))
33                 return NULL;
34
35         pud = pud_offset(pgd, addr);
36         if (pud_none_or_clear_bad(pud))
37                 return NULL;
38
39         pmd = pmd_offset(pud, addr);
40         if (pmd_none_or_clear_bad(pmd))
41                 return NULL;
42
43         return pmd;
44 }
45
46 static pmd_t *alloc_new_pmd(struct mm_struct *mm, unsigned long addr)
47 {
48         pgd_t *pgd;
49         pud_t *pud;
50         pmd_t *pmd;
51
52         pgd = pgd_offset(mm, addr);
53         pud = pud_alloc(mm, pgd, addr);
54         if (!pud)
55                 return NULL;
56
57         pmd = pmd_alloc(mm, pud, addr);
58         if (!pmd)
59                 return NULL;
60
61         if (!pmd_present(*pmd) && __pte_alloc(mm, pmd, addr))
62                 return NULL;
63
64         return pmd;
65 }
66
67 static void move_ptes(struct vm_area_struct *vma, pmd_t *old_pmd,
68                 unsigned long old_addr, unsigned long old_end,
69                 struct vm_area_struct *new_vma, pmd_t *new_pmd,
70                 unsigned long new_addr)
71 {
72         struct address_space *mapping = NULL;
73         struct mm_struct *mm = vma->vm_mm;
74         pte_t *old_pte, *new_pte, pte;
75         spinlock_t *old_ptl, *new_ptl;
76
77         if (vma->vm_file) {
78                 /*
79                  * Subtle point from Rajesh Venkatasubramanian: before
80                  * moving file-based ptes, we must lock vmtruncate out,
81                  * since it might clean the dst vma before the src vma,
82                  * and we propagate stale pages into the dst afterward.
83                  */
84                 mapping = vma->vm_file->f_mapping;
85                 spin_lock(&mapping->i_mmap_lock);
86                 if (new_vma->vm_truncate_count &&
87                     new_vma->vm_truncate_count != vma->vm_truncate_count)
88                         new_vma->vm_truncate_count = 0;
89         }
90
91         /*
92          * We don't have to worry about the ordering of src and dst
93          * pte locks because exclusive mmap_sem prevents deadlock.
94          */
95         old_pte = pte_offset_map_lock(mm, old_pmd, old_addr, &old_ptl);
96         new_pte = pte_offset_map_nested(new_pmd, new_addr);
97         new_ptl = pte_lockptr(mm, new_pmd);
98         if (new_ptl != old_ptl)
99                 spin_lock(new_ptl);
100
101         for (; old_addr < old_end; old_pte++, old_addr += PAGE_SIZE,
102                                    new_pte++, new_addr += PAGE_SIZE) {
103                 if (pte_none(*old_pte))
104                         continue;
105                 pte = ptep_clear_flush(vma, old_addr, old_pte);
106                 /* ZERO_PAGE can be dependant on virtual addr */
107                 pte = move_pte(pte, new_vma->vm_page_prot, old_addr, new_addr);
108                 set_pte_at(mm, new_addr, new_pte, pte);
109         }
110
111         if (new_ptl != old_ptl)
112                 spin_unlock(new_ptl);
113         pte_unmap_nested(new_pte - 1);
114         pte_unmap_unlock(old_pte - 1, old_ptl);
115         if (mapping)
116                 spin_unlock(&mapping->i_mmap_lock);
117 }
118
119 #define LATENCY_LIMIT   (64 * PAGE_SIZE)
120
121 static unsigned long move_page_tables(struct vm_area_struct *vma,
122                 unsigned long old_addr, struct vm_area_struct *new_vma,
123                 unsigned long new_addr, unsigned long len)
124 {
125         unsigned long extent, next, old_end;
126         pmd_t *old_pmd, *new_pmd;
127
128         old_end = old_addr + len;
129         flush_cache_range(vma, old_addr, old_end);
130
131         for (; old_addr < old_end; old_addr += extent, new_addr += extent) {
132                 cond_resched();
133                 next = (old_addr + PMD_SIZE) & PMD_MASK;
134                 if (next - 1 > old_end)
135                         next = old_end;
136                 extent = next - old_addr;
137                 old_pmd = get_old_pmd(vma->vm_mm, old_addr);
138                 if (!old_pmd)
139                         continue;
140                 new_pmd = alloc_new_pmd(vma->vm_mm, new_addr);
141                 if (!new_pmd)
142                         break;
143                 next = (new_addr + PMD_SIZE) & PMD_MASK;
144                 if (extent > next - new_addr)
145                         extent = next - new_addr;
146                 if (extent > LATENCY_LIMIT)
147                         extent = LATENCY_LIMIT;
148                 move_ptes(vma, old_pmd, old_addr, old_addr + extent,
149                                 new_vma, new_pmd, new_addr);
150         }
151
152         return len + old_addr - old_end;        /* how much done */
153 }
154
155 static unsigned long move_vma(struct vm_area_struct *vma,
156                 unsigned long old_addr, unsigned long old_len,
157                 unsigned long new_len, unsigned long new_addr)
158 {
159         struct mm_struct *mm = vma->vm_mm;
160         struct vm_area_struct *new_vma;
161         unsigned long vm_flags = vma->vm_flags;
162         unsigned long new_pgoff;
163         unsigned long moved_len;
164         unsigned long excess = 0;
165         unsigned long hiwater_vm;
166         int split = 0;
167
168         /*
169          * We'd prefer to avoid failure later on in do_munmap:
170          * which may split one vma into three before unmapping.
171          */
172         if (mm->map_count >= sysctl_max_map_count - 3)
173                 return -ENOMEM;
174
175         new_pgoff = vma->vm_pgoff + ((old_addr - vma->vm_start) >> PAGE_SHIFT);
176         new_vma = copy_vma(&vma, new_addr, new_len, new_pgoff);
177         if (!new_vma)
178                 return -ENOMEM;
179
180         moved_len = move_page_tables(vma, old_addr, new_vma, new_addr, old_len);
181         if (moved_len < old_len) {
182                 /*
183                  * On error, move entries back from new area to old,
184                  * which will succeed since page tables still there,
185                  * and then proceed to unmap new area instead of old.
186                  */
187                 move_page_tables(new_vma, new_addr, vma, old_addr, moved_len);
188                 vma = new_vma;
189                 old_len = new_len;
190                 old_addr = new_addr;
191                 new_addr = -ENOMEM;
192         }
193
194         /* Conceal VM_ACCOUNT so old reservation is not undone */
195         if (vm_flags & VM_ACCOUNT) {
196                 vma->vm_flags &= ~VM_ACCOUNT;
197                 excess = vma->vm_end - vma->vm_start - old_len;
198                 if (old_addr > vma->vm_start &&
199                     old_addr + old_len < vma->vm_end)
200                         split = 1;
201         }
202
203         /*
204          * If we failed to move page tables we still do total_vm increment
205          * since do_munmap() will decrement it by old_len == new_len.
206          *
207          * Since total_vm is about to be raised artificially high for a
208          * moment, we need to restore high watermark afterwards: if stats
209          * are taken meanwhile, total_vm and hiwater_vm appear too high.
210          * If this were a serious issue, we'd add a flag to do_munmap().
211          */
212         hiwater_vm = mm->hiwater_vm;
213         mm->total_vm += new_len >> PAGE_SHIFT;
214         vm_stat_account(mm, vma->vm_flags, vma->vm_file, new_len>>PAGE_SHIFT);
215
216         if (do_munmap(mm, old_addr, old_len) < 0) {
217                 /* OOM: unable to split vma, just get accounts right */
218                 vm_unacct_memory(excess >> PAGE_SHIFT);
219                 excess = 0;
220         }
221         mm->hiwater_vm = hiwater_vm;
222
223         /* Restore VM_ACCOUNT if one or two pieces of vma left */
224         if (excess) {
225                 vma->vm_flags |= VM_ACCOUNT;
226                 if (split)
227                         vma->vm_next->vm_flags |= VM_ACCOUNT;
228         }
229
230         if (vm_flags & VM_LOCKED) {
231                 mm->locked_vm += new_len >> PAGE_SHIFT;
232                 if (new_len > old_len)
233                         make_pages_present(new_addr + old_len,
234                                            new_addr + new_len);
235         }
236
237         return new_addr;
238 }
239
240 /*
241  * Expand (or shrink) an existing mapping, potentially moving it at the
242  * same time (controlled by the MREMAP_MAYMOVE flag and available VM space)
243  *
244  * MREMAP_FIXED option added 5-Dec-1999 by Benjamin LaHaise
245  * This option implies MREMAP_MAYMOVE.
246  */
247 unsigned long do_mremap(unsigned long addr,
248         unsigned long old_len, unsigned long new_len,
249         unsigned long flags, unsigned long new_addr)
250 {
251         struct mm_struct *mm = current->mm;
252         struct vm_area_struct *vma;
253         unsigned long ret = -EINVAL;
254         unsigned long charged = 0;
255
256         if (flags & ~(MREMAP_FIXED | MREMAP_MAYMOVE))
257                 goto out;
258
259         if (addr & ~PAGE_MASK)
260                 goto out;
261
262         old_len = PAGE_ALIGN(old_len);
263         new_len = PAGE_ALIGN(new_len);
264
265         /*
266          * We allow a zero old-len as a special case
267          * for DOS-emu "duplicate shm area" thing. But
268          * a zero new-len is nonsensical.
269          */
270         if (!new_len)
271                 goto out;
272
273         /* new_addr is only valid if MREMAP_FIXED is specified */
274         if (flags & MREMAP_FIXED) {
275                 if (new_addr & ~PAGE_MASK)
276                         goto out;
277                 if (!(flags & MREMAP_MAYMOVE))
278                         goto out;
279
280                 if (new_len > TASK_SIZE || new_addr > TASK_SIZE - new_len)
281                         goto out;
282
283                 /* Check if the location we're moving into overlaps the
284                  * old location at all, and fail if it does.
285                  */
286                 if ((new_addr <= addr) && (new_addr+new_len) > addr)
287                         goto out;
288
289                 if ((addr <= new_addr) && (addr+old_len) > new_addr)
290                         goto out;
291
292                 ret = do_munmap(mm, new_addr, new_len);
293                 if (ret)
294                         goto out;
295         }
296
297         /*
298          * Always allow a shrinking remap: that just unmaps
299          * the unnecessary pages..
300          * do_munmap does all the needed commit accounting
301          */
302         if (old_len >= new_len) {
303                 ret = do_munmap(mm, addr+new_len, old_len - new_len);
304                 if (ret && old_len != new_len)
305                         goto out;
306                 ret = addr;
307                 if (!(flags & MREMAP_FIXED) || (new_addr == addr))
308                         goto out;
309                 old_len = new_len;
310         }
311
312         /*
313          * Ok, we need to grow..  or relocate.
314          */
315         ret = -EFAULT;
316         vma = find_vma(mm, addr);
317         if (!vma || vma->vm_start > addr)
318                 goto out;
319         if (is_vm_hugetlb_page(vma)) {
320                 ret = -EINVAL;
321                 goto out;
322         }
323         /* We can't remap across vm area boundaries */
324         if (old_len > vma->vm_end - addr)
325                 goto out;
326         if (vma->vm_flags & (VM_DONTEXPAND | VM_PFNMAP)) {
327                 if (new_len > old_len)
328                         goto out;
329         }
330         if (vma->vm_flags & VM_LOCKED) {
331                 unsigned long locked, lock_limit;
332                 locked = mm->locked_vm << PAGE_SHIFT;
333                 lock_limit = current->signal->rlim[RLIMIT_MEMLOCK].rlim_cur;
334                 locked += new_len - old_len;
335                 ret = -EAGAIN;
336                 if (locked > lock_limit && !capable(CAP_IPC_LOCK))
337                         goto out;
338         }
339         if (!may_expand_vm(mm, (new_len - old_len) >> PAGE_SHIFT)) {
340                 ret = -ENOMEM;
341                 goto out;
342         }
343
344         if (vma->vm_flags & VM_ACCOUNT) {
345                 charged = (new_len - old_len) >> PAGE_SHIFT;
346                 if (security_vm_enough_memory(charged))
347                         goto out_nc;
348         }
349
350         /* old_len exactly to the end of the area..
351          * And we're not relocating the area.
352          */
353         if (old_len == vma->vm_end - addr &&
354             !((flags & MREMAP_FIXED) && (addr != new_addr)) &&
355             (old_len != new_len || !(flags & MREMAP_MAYMOVE))) {
356                 unsigned long max_addr = TASK_SIZE;
357                 if (vma->vm_next)
358                         max_addr = vma->vm_next->vm_start;
359                 /* can we just expand the current mapping? */
360                 if (max_addr - addr >= new_len) {
361                         int pages = (new_len - old_len) >> PAGE_SHIFT;
362
363                         vma_adjust(vma, vma->vm_start,
364                                 addr + new_len, vma->vm_pgoff, NULL);
365
366                         mm->total_vm += pages;
367                         vm_stat_account(mm, vma->vm_flags, vma->vm_file, pages);
368                         if (vma->vm_flags & VM_LOCKED) {
369                                 mm->locked_vm += pages;
370                                 make_pages_present(addr + old_len,
371                                                    addr + new_len);
372                         }
373                         ret = addr;
374                         goto out;
375                 }
376         }
377
378         /*
379          * We weren't able to just expand or shrink the area,
380          * we need to create a new one and move it..
381          */
382         ret = -ENOMEM;
383         if (flags & MREMAP_MAYMOVE) {
384                 if (!(flags & MREMAP_FIXED)) {
385                         unsigned long map_flags = 0;
386                         if (vma->vm_flags & VM_MAYSHARE)
387                                 map_flags |= MAP_SHARED;
388
389                         new_addr = get_unmapped_area(vma->vm_file, 0, new_len,
390                                                 vma->vm_pgoff, map_flags);
391                         ret = new_addr;
392                         if (new_addr & ~PAGE_MASK)
393                                 goto out;
394                 }
395                 ret = move_vma(vma, addr, old_len, new_len, new_addr);
396         }
397 out:
398         if (ret & ~PAGE_MASK)
399                 vm_unacct_memory(charged);
400 out_nc:
401         return ret;
402 }
403
404 asmlinkage unsigned long sys_mremap(unsigned long addr,
405         unsigned long old_len, unsigned long new_len,
406         unsigned long flags, unsigned long new_addr)
407 {
408         unsigned long ret;
409
410         down_write(&current->mm->mmap_sem);
411         ret = do_mremap(addr, old_len, new_len, flags, new_addr);
412         up_write(&current->mm->mmap_sem);
413         return ret;
414 }